nvme-rdma: use ib_client API to detect device removal
authorSteve Wise <swise@opengridcomputing.com>
Fri, 2 Sep 2016 16:01:54 +0000 (09:01 -0700)
committerSagi Grimberg <sagi@grimberg.me>
Mon, 12 Sep 2016 19:29:41 +0000 (22:29 +0300)
Change nvme-rdma to use the IB Client API to detect device removal.
This has the wonderful benefit of being able to blow away all the
ib/rdma_cm resources for the device being removed.  No craziness about
not destroying the cm_id handling the event.  No deadlocks due to broken
iw_cm/rdma_cm/iwarp dependencies.  And no need to have a bound cm_id
around during controller recovery/reconnect to catch device removal
events.

We don't use the device_add aspect of the ib_client service since we only
want to create resources for an IB device if we have a target utilizing
that device.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
drivers/nvme/host/rdma.c

index eeb08b6..d6bdf55 100644 (file)
@@ -1320,64 +1320,6 @@ out_destroy_queue_ib:
        return ret;
 }
 
-/**
- * nvme_rdma_device_unplug() - Handle RDMA device unplug
- * @queue:      Queue that owns the cm_id that caught the event
- *
- * DEVICE_REMOVAL event notifies us that the RDMA device is about
- * to unplug so we should take care of destroying our RDMA resources.
- * This event will be generated for each allocated cm_id.
- *
- * In our case, the RDMA resources are managed per controller and not
- * only per queue. So the way we handle this is we trigger an implicit
- * controller deletion upon the first DEVICE_REMOVAL event we see, and
- * hold the event inflight until the controller deletion is completed.
- *
- * One exception that we need to handle is the destruction of the cm_id
- * that caught the event. Since we hold the callout until the controller
- * deletion is completed, we'll deadlock if the controller deletion will
- * call rdma_destroy_id on this queue's cm_id. Thus, we claim ownership
- * of destroying this queue before-hand, destroy the queue resources,
- * then queue the controller deletion which won't destroy this queue and
- * we destroy the cm_id implicitely by returning a non-zero rc to the callout.
- */
-static int nvme_rdma_device_unplug(struct nvme_rdma_queue *queue)
-{
-       struct nvme_rdma_ctrl *ctrl = queue->ctrl;
-       int ret = 0;
-
-       /* Own the controller deletion */
-       if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
-               return 0;
-
-       dev_warn(ctrl->ctrl.device,
-               "Got rdma device removal event, deleting ctrl\n");
-
-       /* Get rid of reconnect work if its running */
-       cancel_delayed_work_sync(&ctrl->reconnect_work);
-
-       /* Disable the queue so ctrl delete won't free it */
-       if (!test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags)) {
-               /* Free this queue ourselves */
-               nvme_rdma_stop_queue(queue);
-               nvme_rdma_destroy_queue_ib(queue);
-
-               /* Return non-zero so the cm_id will destroy implicitly */
-               ret = 1;
-       }
-
-       /*
-        * Queue controller deletion. Keep a reference until all
-        * work is flushed since delete_work will free the ctrl mem
-        */
-       kref_get(&ctrl->ctrl.kref);
-       queue_work(nvme_rdma_wq, &ctrl->delete_work);
-       flush_work(&ctrl->delete_work);
-       nvme_put_ctrl(&ctrl->ctrl);
-
-       return ret;
-}
-
 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
                struct rdma_cm_event *ev)
 {
@@ -1419,8 +1361,8 @@ static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
                nvme_rdma_error_recovery(queue->ctrl);
                break;
        case RDMA_CM_EVENT_DEVICE_REMOVAL:
-               /* return 1 means impliciy CM ID destroy */
-               return nvme_rdma_device_unplug(queue);
+               /* device removal is handled via the ib_client API */
+               break;
        default:
                dev_err(queue->ctrl->ctrl.device,
                        "Unexpected RDMA CM event (%d)\n", ev->event);
@@ -2030,27 +1972,57 @@ static struct nvmf_transport_ops nvme_rdma_transport = {
        .create_ctrl    = nvme_rdma_create_ctrl,
 };
 
+static void nvme_rdma_add_one(struct ib_device *ib_device)
+{
+}
+
+static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
+{
+       struct nvme_rdma_ctrl *ctrl;
+
+       /* Delete all controllers using this device */
+       mutex_lock(&nvme_rdma_ctrl_mutex);
+       list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
+               if (ctrl->device->dev != ib_device)
+                       continue;
+               dev_info(ctrl->ctrl.device,
+                       "Removing ctrl: NQN \"%s\", addr %pISp\n",
+                       ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
+               __nvme_rdma_del_ctrl(ctrl);
+       }
+       mutex_unlock(&nvme_rdma_ctrl_mutex);
+
+       flush_workqueue(nvme_rdma_wq);
+}
+
+static struct ib_client nvme_rdma_ib_client = {
+       .name   = "nvme_rdma",
+       .add = nvme_rdma_add_one,
+       .remove = nvme_rdma_remove_one
+};
+
 static int __init nvme_rdma_init_module(void)
 {
+       int ret;
+
        nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
        if (!nvme_rdma_wq)
                return -ENOMEM;
 
+       ret = ib_register_client(&nvme_rdma_ib_client);
+       if (ret) {
+               destroy_workqueue(nvme_rdma_wq);
+               return ret;
+       }
+
        nvmf_register_transport(&nvme_rdma_transport);
        return 0;
 }
 
 static void __exit nvme_rdma_cleanup_module(void)
 {
-       struct nvme_rdma_ctrl *ctrl;
-
        nvmf_unregister_transport(&nvme_rdma_transport);
-
-       mutex_lock(&nvme_rdma_ctrl_mutex);
-       list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
-               __nvme_rdma_del_ctrl(ctrl);
-       mutex_unlock(&nvme_rdma_ctrl_mutex);
-
+       ib_unregister_client(&nvme_rdma_ib_client);
        destroy_workqueue(nvme_rdma_wq);
 }